Know if the deal is good before you make the offer.

The engine is a dependency-free TypeScript module with no knowledge of the DOM, the server or the database, so the same code runs in the browser for the sliders and on the server for PDFs. Mortgage, expenses, ratios, projections and color thresholds are separate files with typed inputs and outputs. IRR is solved with Newton-Raphson and falls back to bisection when the cash-flow series misbehaves. Before any screen was built, the engine was validated against the broker’s own underwriting cases; ten of eleven figures match exactly and the eleventh is within three cents. That validation script still runs, so a formula change that drifts fails before it ships.

• Zero-dependency module, runs client and server
• Newton-Raphson IRR with bisection fallback
• Validated against the broker’s underwriting, script kept in the repo
• Every input change recomputes through useMemo

Five AI tools ship in the app: rent estimation, zip-code price and rent growth research, an investor chatbot that can see the deal on screen, a Property Scout that ranks candidates against a chosen strategy, and a neighborhood scorecard. All of them run Claude Haiku 4.5 through Vercel AI Gateway, authenticated with OIDC so production holds no static AI key. Rent estimation tries the RentCast API first and only falls back to the model; growth research is grounded in embedded Texas market data for twenty-five plus zip codes. No financial arithmetic ever goes to a model.

• Claude Haiku 4.5 via Vercel AI Gateway, OIDC auth
• RentCast comps first, AI as the fallback
• Embedded Texas market reference data as grounding
• Deterministic math never leaves the engine

A RESO Web API client talks to the broker’s MLS through its back-office feed. Credentials exchange for a bearer token that is cached in Postgres and refreshed an hour before expiry. Active duplex listings render server-side on a listings page for logged-in users, every listing hands its MLS number to the analyzer for a pre-filled analysis, and the Scout searches the same inventory. The MLS reviewed the actual flows and confirmed the paid-tool model before anything went live.

• RESO Web API / OData client with cached bearer token
• Server-rendered listings page behind login
• Analyzer pre-fill from an MLS number in the URL
• Usage reviewed and approved by the MLS

Next.js 16 App Router on React 19 with three route groups: public marketing, minimal-chrome auth, and the app shell. Supabase Postgres holds profiles, saved deals, folders, comparisons and thresholds, with row-level security on every user table and a trigger that creates the profile on signup. Stripe checkout and a webhook keep each profile’s plan in sync. Admins manage thresholds, users and a usage log. The consultation form runs behind BotID and a honeypot, delivers by Resend, and mirrors every lead into the broker’s CRM.

• Supabase Auth + Postgres, RLS on every user table
• Stripe plans, portal and webhook plan sync
• Admin panel: thresholds, users, usage log
• Resend delivery with a CRM mirror, BotID-protected

A property-records API was integrated, evaluated, and removed. At $500 a month for five thousand calls it auto-filled exactly two fields, both of which the broker types in anyway. Two other data vendors were passed over for unpredictable per-call fees, and county-record scraping was shelved because every Texas county runs a different system. The all-in infrastructure bill for the shipped product is under $100 a month.

Cap rate, DSCR, IRR and every other figure are deterministic code, and the AI is confined to the questions that genuinely need judgement: what would this rent for, what has this zip code done, what should a first-time buyer make of a yellow ratio. That keeps the answers reproducible, keeps the AI bill in the tens of dollars, and means a model retirement cannot change a client’s numbers.

A week before go-live the broker asked to pull public pricing, self-serve signup and the tools pages from the site until he had a way to protect his methodology from copycats. The routes stayed, the nav changed, the retired pages redirect for visitors and still work for signed-in users, and he onboards agents by sharing a direct link. The public site now sells a conversation, and the software is what his clients get after it.

TREC broker identification, the IABS form, the Consumer Protection Notice and Equal Housing language live on a first-class disclosures page and beside every tool that produces a number. The broker’s home address was removed from the footer, legal pages and structured data when he flagged it, and a comment in the source says not to add it back.